Antarctic ice shelf crevasse detection using multi-source remote sensing data and machine learning

Ice crevasses are pervasive features across the Arctic and Antarctic ice sheets. These deep, open fractures in the ice surface serve as critical conduits for transporting surface meltwater into the englacial system, significantly impacting ice sheet hydrology and stability. Accurate mapping of the spatial and temporal distribution of ice crevasses is vital for advancing our understanding of ice sheet dynamics and their evolution. Remote sensing technology provides a robust platform to achieve this purpose, while the rapid advancement of machine learning algorithms offers substantial benefits for automated crevasse detection, facilitating efficient and large-scale mapping. This study conducts a comprehensive comparison of the performance of various machine learning models, including CNN, U-Net, ResNet, and DeepLab, for ice crevasse extraction. Through quantitative evaluation metrics and visual inspection, the optimal machine learning model was selected to map ice crevasses on Antarctic ice shelves using multi-source remote sensing data, such as SAR and optical satellite imagery. Furthermore, this work explores the strengths and limitations of various machine learning in detecting ice crevasse and proposes potential solutions for further refinement. This study aims to contributes to enhancing ice crevasse detection and offering robust ice crevasse datasets, which is crucial for reliable analyzing the dynamic of the Antarctic ice sheet in the future.